How accurate can data fusion be?

Abstract

"In this paper we structure the validity a data fusion procedure may achieve by four levels. It is shown that the first level is meaningless and only the last, fourth, level typically is controlled when traditional techniques of statistical matching are applied. The preservation of the joint distribution and the correlation structure of the variables not jointly observed can be evaluated by using the non-iterative multiple imputation procedure NIBAS. In a simulation study, we find the multiple imputation approaches superior to the traditional matching techniques. Auxiliary data can easily and efficiently be used by standard MI procedures such as NORM provided by Schafer (1997). Finally, to avoid identification problems we suggest to apply split questionnaire survey (SQS) designs as proposed by Raghunathan and Grizzle (1995). For the SQS design special patterns of missingness are created. The missing data can then quite successfully be multiply imputed." (Author's abstract, IAB-Doku) ((en))

Cite article

Rässler, S. (2003): How accurate can data fusion be? Erlangen u.a., 4 p.